China's AI Giants Ignite Price War: Alibaba and DeepSeek Slash Model Costs

The artificial intelligence landscape is witnessing a heightened competitive environment, marked by the introduction of new, powerful models from major Chinese tech firms. Alibaba has unveiled Qwen3.8-Max, its most extensive AI model to date, while DeepSeek's latest V4-Flash model is attracting significant attention for its remarkably low inference pricing, challenging established systems.

Alibaba's Qwen3.8-Max boasts an impressive 2.4 trillion parameters and utilizes a mixture-of-experts (MoE) architecture. This design intelligently activates only a subset of the model for each request, specifically around 95 billion parameters at any given time, which Alibaba claims significantly reduces operational costs and response latency compared to activating the full model. Qwen3.8-Max is a multimodal powerhouse, capable of processing text, images, and video, and supports an extensive context window of up to one million tokens. Alibaba has demonstrated its capabilities, noting that the model successfully completed a software engineering project over 16 days. Its sheer scale positions it closely to Moonshot AI’s Kimi K3, which features 2.8 trillion total parameters with approximately 104 billion active during inference. Competition extends to pricing, with Qwen3.8-Max costing $2 per million input tokens and $6 per million output tokens, a notable difference from Kimi K3's $3 and $15, respectively, for similar token usage. Following its release, Qwen3.8-Max ascended to the top position among Chinese text models on Arena.AI, a crowdsourced comparison platform, though it remained behind certain Anthropic models in the overall rankings. It also secured second place on Arena.AI’s leaderboard for image and visual material analysis, trailing only an Anthropic Claude Fable 5 variant.

DeepSeek, in contrast, has opted for a distinct strategy with its V4-Flash model, prioritizing aggressive inference pricing rather than matching the colossal scale of Alibaba’s and Moonshot AI’s newest offerings. Artificial Analysis reports V4-Flash’s costs at a mere $0.14 per million input tokens and $0.28 per million output tokens, making it considerably more affordable than several widely utilized AI systems. This model features a one-million-token context window and 284 billion total parameters, with 13 billion active during inference, employing a sparse architecture similar to Qwen3.8-Max but at a smaller scale. DeepSeek further innovates with cache-hit pricing for the Max Effort version of V4-Flash, listed at $0.003 per million tokens, an extraordinary 98% reduction from its standard input rate for reusing previously processed context. Artificial Analysis' benchmark testing confirmed DeepSeek’s cost efficiency, estimating V4-Flash’s average cost at three cents per test, a stark contrast to 86 cents for Kimi K3, $1.86 for OpenAI’s GPT-5.6 Sol, and $3.15 for Anthropic’s Claude Fable 5. This comparison factors in the input and output volumes each model consumes for benchmark completion. DeepSeek V4-Flash (Max Effort reasoning version) achieved a score of 40 on Artificial Analysis’ Intelligence Index and demonstrated an output rate of approximately 118 tokens per second during testing.

The comparison between these models underscores that model size alone does not dictate inference cost; factors such as architecture, active parameter count, token consumption, and the number of calls required for task completion significantly influence the total operational cost. Moonshot AI’s Kimi K3 serves as another prime example of how advertised API prices can diverge from the actual cost of longer, more complex workloads. Artificial Analysis lists Kimi K3 at $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million tokens. However, on Artificial Analysis’ AA-Briefcase benchmark for agentic knowledge work, Kimi K3 averaged $10.57 per task, generating around 120,000 output tokens and requiring an average of 83 turns per task. This cost reflects Kimi K3’s token pricing, substantial output volume, and numerous model interactions. Consequently, repeated model calls and larger outputs can escalate the total cost beyond initial per-token rates. Kimi K3 achieved the second-highest overall score on the AA-Briefcase evaluation at the time of testing, only behind Claude Fable 5, and scored 57 on Artificial Analysis’ broader Intelligence Index. These cost-per-task measurements offer crucial context to standard API pricing, revealing how models with varied architectures and usage patterns can consume vastly different amounts of compute and tokens for similar tasks.

Another influential aspect shaping deployment options and costs is the distribution strategy adopted by Chinese developers. Alibaba, DeepSeek, and Moonshot AI are actively supporting open-weight releases alongside their hosted API access, granting developers greater flexibility. DeepSeek V4-Flash is an open-weight model under an MIT license, with weights accessible via Hugging Face. Similarly, Kimi K3 is available as an open-weight model under Moonshot AI’s proprietary license. This open-weight approach allows developers to run models on their own infrastructure or through third-party providers, decoupling access from reliance on a single developer-hosted inference service. While deployment costs still depend on the chosen hardware and infrastructure, the model itself is not confined to a single hosted API. This contrasts with leading models from OpenAI, Anthropic, and Google, which typically maintain closed model weights. Lian Jye Su, chief analyst at Omdia, highlights the significance of this trend, stating that many business workflows prioritize models that are